Selection Bias In Estimates Research Articles

Comparison of Propensity Score Weighting Methods to Remove Selection Bias in Average Treatment Effect Estimates

In this Monte Carlo simulation study, the performance of six different propensity score methods implemented through weighting cases was investigated: inverse probability of treatment weighting, truncated inverse probability of treatment weighting, propensity score stratification, marginal mean weighting through propensity score stratification, optimal full propensity score matching, and marginal mean weighting through optimal full propensity score matching. These methods aim to reduce selection bias in estimates of the average treatment effect (ATE) in observational studies. For the estimation of standard errors of the ATE with weights, three methods were compared: weighted least squares (WLS), Taylor series linearization (TSL), and jackknife (JK). Results indicated that covariance adjustment extensions of the investigated propensity score methods, in combination with TSL and JK standard error estimation methods, remove the selection bias appropriately and provide the most accurate standard errors under the simulated conditions.

What predicts willingness to participate in a follow-up panel study among respondents to a national web/mail survey?

The American Family Health Study (AFHS) collected family health and fertility data from a national probability sample of persons aged 18-49 between September 2021 and May 2022, using web and mail exclusively. In July 2022, we surveyed AFHS respondents and gauged their willingness to become part of a national web panel that would create novel longitudinal data on these topics. We focus on predictors of willingness to participate, identifying the potential selection bias that this type of approach may introduce. We found that efforts of this type to create a national web panel may introduce potential selection bias in estimates based on the panel respondents, with individuals having higher socio-economic status being more cooperative. Thus, alternative recruitment strategies and re-weighting of the subsample may be needed to further reduce selection bias. We present methodological implications of our results, limitations of our approach, and suggestions for further research on this topic.

Evaluating Pre-election Polling Estimates Using a New Measure of Non-ignorable Selection Bias.

Among the numerous explanations that have been offered for recent errors in pre-election polls, selection bias due to non-ignorable partisan nonresponse bias, where the probability of responding to a poll is a function of the candidate preference that a poll is attempting to measure (even after conditioning on other relevant covariates used for weighting adjustments), has received relatively less focus in the academic literature. Under this type of selection mechanism, estimates of candidate preferences based on individual or aggregated polls may be subject to significant bias, even after standard weighting adjustments. Until recently, methods for measuring and adjusting for this type of non-ignorable selection bias have been unavailable. Fortunately, recent developments in the methodological literature have provided political researchers with easy-to-use measures of non-ignorable selection bias. In this study, we apply a new measure that has been developed specifically for estimated proportions to this challenging problem. We analyze data from 18 different pre-election polls: 9 different telephone polls conducted in 8 different states prior to the US presidential election in 2020, and nine different pre-election polls conducted either online or via telephone in Great Britain prior to the 2015 general election. We rigorously evaluate the ability of this new measure to detect and adjust for selection bias in estimates of the proportion of likely voters that will vote for a specific candidate, using official outcomes from each election as benchmarks and alternative data sources for estimating key characteristics of the likely voter populations in each context.

Testing for selection bias and moral hazard in private health insurance: Evidence from a mixed public-private health system.

Separating selection bias from moral hazard in private health insurance (PHI) markets has been a challenging task. We estimate selection bias and moral hazard in Australia's mixed public-private health system, where PHI premiums are community-rated rather than risk-rated. Using longitudinal cohort data, with fine-grained measures for medical services predominantly funded by PHI providers, we find consistent and robust estimates of advantageous selection among hospitalized cardiovascular disease (CVD) patients. Specifically, we show that in addition to their risk-averse attributes, CVD patients who purchase PHI use fewer services that are not covered by PHI providers (e.g., general practitioners and emergency departments) and have fewer comorbidities. Finally, unlike previous studies, we show that ex-post moral hazard exists in the use of specific "in-hospital" medical services such as specialist and physician services, miscellaneous diagnostic procedures, and therapeutic treatments. From the perspective of PHI providers, the annual cost of moral hazard translates to a lower bound estimate of $707 per patient, equivalent to a 3.03% reduction in their annual profits.

Trends in long-term cancer survival in Cali, Colombia: 1998-2017.

Population-based cancer survival is an indicator of the effectiveness of cancer services that reflects the survival of all cancer patients in the population, regardless of socioeconomic status and disease characteristics. Provision of an up-to-date survival estimate of patients recorded within Cali Population Cancer Registry (RPCC) in 1998-2017. As a second objective, results will be compared with those reported by the CONCORD study for cancers prioritized by the current Ten-Year Cancer Control Plan of Colombia, 2012-2021. Adult cancer cases (aged 15 to 99 years) for nine cancer types diagnosed between 1998 and 2017, with follow-up to 2018, were obtained from the RPCC. The 5-year age-standardized net survival estimates (NS) were estimated using the Pohar-Perme. The results for the period 1995- 2014 were compared with those reported by the CONCORD study for the following locations: stomach (C16), breast (C50), cervix (C53), prostate (C61), and lung (C33-34). Five-year survival estimates for breast and prostate cancers improved ten percentage points through 2007 (70.8 to 81.1 for breast and 79.9 to 90.2 for prostate) and remained stable during 2008-2017. For cervical cancer, survival estimates has remained stable for the last two decades at 53%. For stomach cancer and lung cancer, five-year NS was lower than 25% over the study period. For colorectal cancer, survival estimates increased from 37.9% in 1998-2002 to 54.8% in 2013-2017. Compared to previous 5-year survival estimates of cases diagnosed in 2010-2014, the estimates in this study are significantly higher than those obtained by CONCORD. Survival estimates of patients diagnosed in 1995-2009 showed no difference to CONCORD study. Periodic update of vital status and date of last contact reduces bias in survival estimates in population-based cancer registries with passive follow-up.

A natural history of gastroenteropancreatic neuroendocrine tumors evaluated at a tertiary cancer center: An observational study of 2,067 cases.

508 Background: Gastroenteropancreatic neuroendocrine tumors (GEP-NETs) are the second most prevalent digestive cancer and are increasing in incidence. With relatively indolent tumor growth, survival is often long, even for patients with metastatic disease. However, few studies have assessed long-term survival. Therefore, we analyzed the overall survival (OS) of a cohort of GEP-NET patients at a tertiary cancer center from 2000-2015. Methods: 2,707 patients with a pathologic diagnosis of GEP-NET were identified. 640 patients presented for second-opinion visits and were excluded, leaving 2,067 patients, of whom, 2,053 had vital status data for analysis. Primary sites included jejunum/ileum (small bowel), pancreas, gastric, colorectum, appendix, duodenum, and unknown. OS time was calculated from the date of pathologic diagnosis to the date of last follow-up or death. The reverse Kaplan-Meier method was used to calculate follow-up time. Vital status was identified through linkage to the Florida Cancer Registry. Kaplan-Meier (KM) curves and log-rank test were used for OS analyses. Results: The median age of the cohort was 60.7 years. The median follow-up time was 148.9 months (12.4 years). 1,431 (69.2%) patients presented with distant metastatic disease. Most patients, 1,351 (65.5%), underwent treatment at another facility before presentation. The median OS was 132.2 months (11.0 years) for the cohort and was 108.0 and 245.5 months for patients presenting with and without metastasis, respectively (p < 0.001). For metastatic patients, unknown primary tumor site was associated with the worst median OS (55 months), followed by rectal (77 months), gastric (84 months) and pancreatic (85 months; p < 0.001). Duodenal and small bowel had the longest median OS at 147 and 143 months, respectively. Surgery for metastatic disease had an OS benefit (155.7 months) compared to no surgery (98.3 months). Conclusions: In a large cohort of predominantly metastatic GEP-NET patients with extended follow-up time, median OS was over 11 years and 9 years for patients with metastatic disease. These results suggest that OS at a tertiary cancer center may be greater than historical national administrative analyses when given adequate follow-up time, which is important to avoid bias in survival estimates.[Table: see text]

Effects of helicopter net‐gunning on survival of bighorn sheep

AbstractWildlife capture, and the data collection associated with it, has led to major advancements in ecology that are integral to decision making pertaining to wildlife conservation. Capturing wildlife, however, can cause lethal and non‐lethal risks to animals. Understanding the factors that contribute to the level of risk involved in wildlife capture is therefore important for the development and implementation of the safest and most effective methodologies. We used data from 736 animal captures of 389 individuals for 2 subspecies of female bighorn sheep (Rocky Mountain bighorn [Ovis canadensis canadensis], Sierra Nevada bighorn sheep [O. c. sierrae]) in Wyoming and California, USA, in 2002–2020 to evaluate the degree and extent of time that capture via helicopter net‐gunning affects survival. We compared pre‐ and post‐capture survival during a 10‐week window centered on a capture event, and post‐capture survival between captured animals and animals that were monitored but not captured during the 10‐week window. Additionally, we evaluated the effects of handling techniques (number of times captured, season of capture event, handling time, chase time, and body temp) and biological factors (age and nutritional condition) on probability of capture mortality. Mean daily survival was 0.9992 during a 5‐week pre‐capture window, dropped to 0.9864 on the day of capture, and rebounded within 3 days of capture to pre‐capture levels and that of sheep that were not captured. Overall, direct mortality resulting from capture was 1.36%, with 0.54% mortality occurring within the 3 days following a capture event for an overall 1.90% capture‐related mortality. The only handling and biological metrics that influenced the probability of capture mortality were rectal temperature and nutritional condition; high initial rectal temperatures and poor body condition were associated with increased risk of mortality in the days following capture. Overall, helicopter net‐gunning imposed low and short‐term risk to survival of female bighorn sheep. To reduce bias in survival estimates, we recommend using a 3‐day censorship window for post‐capture mortalities as opposed to the common practice of a 2–5‐week censor window. Helicopter net‐gunning, including annual or seasonal recaptures, remains an effective and comparatively safe technique for capture and associated data collection of bighorn sheep.

The impacts of marking on bats: mark-recapture models for assessing injury rates and tag loss

AbstractThe individual identification of animals is essential for long-term population ecology studies. Injuries and tag loss can negatively impact animal welfare and parameter estimates. We used time-to-failure curves and multistate mark-recapture models to estimate injury rates, body condition decline, tag loss and survival probabilities associated with two tags commonly used in bats: metallic arm bands and steel ball chain collars. We monitored two colonies of Carollia perspicillata at União Biological Reserve, southeastern Brazil, from 2013 to 2019. Every other month, we marked individuals with arm bands, collars or both. Collars did not cause noticeable injuries, but were lost at a higher rate than arm bands (1.5–2.5 times). The arm band-related injury probability between captures was ~ 0.2, and after 17 months half the marked animals developed injuries. Animals marked as juveniles were less likely to get injured than adults. Injuries were associated with a lower body condition in females. Body condition was positively associated with apparent survival; however, a direct effect of arm band injuries on survival estimates was not observed. The tag loss rates caused a negative bias in survival estimates. Mark-recapture studies should evaluate the efficacy of the marking technique, accounting for potential animal injury, tag loss, and bias in parameter estimation. Marking individuals should be limited to studies that monitor species with high recapture probability, using the least harmful tags possible, and balancing the burden to the animals with the potential to generate knowledge.

Impact of comorbidity assessment methods to predict non-cancer mortality risk in cancer patients: a retrospective observational study using the National Health Insurance Service claims-based data in Korea

BackgroundCancer patients’ prognoses are complicated by comorbidities. Prognostic prediction models with inappropriate comorbidity adjustments yield biased survival estimates. However, an appropriate claims-based comorbidity risk assessment method remains unclear. This study aimed to compare methods used to capture comorbidities from claims data and predict non-cancer mortality risks among cancer patients.MethodsData were obtained from the National Health Insurance Service-National Sample Cohort database in Korea; 2979 cancer patients diagnosed in 2006 were considered. Claims-based Charlson Comorbidity Index was evaluated according to the various assessment methods: different periods in washout window, lookback, and claim types. The prevalence of comorbidities and associated non-cancer mortality risks were compared. The Cox proportional hazards models considering left-truncation were used to estimate the non-cancer mortality risks.ResultsThe prevalence of peptic ulcer, the most common comorbidity, ranged from 1.5 to 31.0%, and the proportion of patients with ≥1 comorbidity ranged from 4.5 to 58.4%, depending on the assessment methods. Outpatient claims captured 96.9% of patients with chronic obstructive pulmonary disease; however, they captured only 65.2% of patients with myocardial infarction. The different assessment methods affected non-cancer mortality risks; for example, the hazard ratios for patients with moderate comorbidity (CCI 3–4) varied from 1.0 (95% CI: 0.6–1.6) to 5.0 (95% CI: 2.7–9.3). Inpatient claims resulted in relatively higher estimates reflective of disease severity.ConclusionsThe prevalence of comorbidities and associated non-cancer mortality risks varied considerably by the assessment methods. Researchers should understand the complexity of comorbidity assessments in claims-based risk assessment and select an optimal approach.

Survival Rates and Capture Heterogeneity of Bottlenose Dolphins (Tursiops truncatus) in the Shannon Estuary, Ireland

Adult survival is arguably the most important demographic parameter for long-lived species as it has a large impact on population growth, and it can be estimated for cetacean populations using natural markings and mark-recapture (MR) modelling. Here we describe a 26-year study of a genetically discrete, resident population of bottlenose dolphins in the Shannon Estuary, Ireland, conducted by an NGO using multiple platforms. We estimated survival rates (SRs) using Cormack-Jolly-Seber models and explored the effects of variable survey effort, multiple researchers, and changes in camera equipment as well as capture heterogeneity induced by changes in marks and site fidelity variation, all common issues affecting longitudinal dolphin studies. The mean adult SR was 0.94 (±0.001 SD) and thus comparable to the estimates reported for other bottlenose dolphin populations. Capture heterogeneity through variation in mark severity was confirmed, with higher capture probabilities for well-marked individuals than for poorly marked individuals and a “transience” effect being detected for less well-marked individuals with 43% only recorded once. Likewise, both SR and capture probabilities were comparatively low for individuals with low site fidelity to the Shannon Estuary, and SR of these individuals additionally decreased even further toward the end of the study, reflecting a terminal bias. This bias was attributed to non-random temporal migration, and, together with high encounter rates in Brandon Bay, supported the hypothesis of range expansion. Our results highlight the importance of consistent and geographically homogenous survey effort and support the differentiation of individuals according to their distinctiveness to avoid biased survival estimates.

Tag Type and Location‐Dependent Retention Impart Varied Levels of Bias on Mark–Recapture Parameter Estimates

AbstractPopulation parameter estimates from mark–recapture studies are dependent on individuals retaining marks or tags. Therefore, tag retention estimates are needed for different tag types and anatomical tagging locations. Few studies have empirically quantified the bias from tag retention on fish population parameters that are derived from mark–recapture studies. We examined differences in retention between T‐bar anchor tags and PIT tags as well as among four anatomical locations for PIT tags in Brown Trout Salmo trutta in a tailwater fishery in Arkansas, USA. We also estimated the relative bias of tag type and PIT tag location on apparent survival estimates from Cormack–Jolly–Seber models. Tag retention for the anchor tags was 15.1% lower than that for the PIT tags after 1 year and 46.1% lower after 4 years. Greater PIT tag retention resulted in less biased estimates of apparent survival for PIT tags (average −7.1%) than for anchor tags (average −37.8%). However, PIT tags that were placed in different anatomical locations had varying retention rates, so the degree of relative bias that was associated with their apparent survival estimates also varied. Inserting the PIT tags in the cheek or dorsal musculature provided the greatest retention for Brown Trout and may provide the least biased apparent survival estimates from future mark–recapture studies.

The impact of excluding or including Death Certificate Initiated (DCI) cases on estimated cancer survival: A simulation study

BackgroundPopulation-based cancer registries strive to cover all cancer cases diagnosed within the population, but some cases will always be missed and no register is 100 % complete. Many cancer registries use death certificates to identify additional cases not captured through other routine sources, to hopefully add a large proportion of the missed cases. Cases notified through this route, who would not have been captured without death certificate information, are referred to as Death Certificate Initiated (DCI) cases. Inclusion of DCI cases in cancer registries increases completeness and is important for estimating cancer incidence. However, inclusion of DCI cases will generally lead to biased estimates of cancer survival, but the same is often also true if excluding DCI cases. Missed cases are probably not a random sample of all cancer cases, but rather cases with poor prognosis. Further, DCI cases have poorer prognosis than missed cases in general, since they have all died with cancer mentioned on the death certificates. MethodsWe performed a simulation study to estimate the impact of including or excluding DCI cases on cancer survival estimates, under different scenarios. ResultsWe demonstrated that including DCI cases underestimates survival. The exclusion of DCI cases gives unbiased survival estimates if missed cases are a random sample of all cancer cases, while survival is overestimated if these have poorer prognosis. ConclusionIn our most extreme scenarios, with 25 % of cases initially missed, the usual practice of including DCI cases underestimated 5-year survival by at most 3 percentage points.

Estimates of lung and pancreatic cancer survival in Sweden with and without inclusion of death certificate initiated (DCI) cases

Introduction International differences in cancer incidence and survival may partly reflect differences in cancer registration practices. As opposed to most other National Cancer Registries, Death Certificate Initiated (DCI) cases are not included in the Swedish Cancer Register. We characterized cases not reported to the Swedish Cancer Register and assessed the impact of inclusion of DCI cases on the completeness and estimates of one-year lung and pancreatic cancer survival. Methods We used information in the Swedish Cause of Death Register to identify individuals in two Health Care Regions (West and Uppsala Örebro) with lung or pancreatic cancer as cause of death in 2013. These records were cross-linked to the Cancer Register to identify individuals without a corresponding cancer registration, i.e. Death Certificate Notified (DCN) cases. DCN cases were cross-linked to the Patient Register to retrieve hospital discharge information to confirm the diagnosis. In a separate step, trace-back of DCN cases was performed to access medical records to validate the diagnosis. Results Following validity checks, an estimated 16% and 34% of individuals with a diagnosis of lung or pancreatic cancer, respectively, had not been reported to the SCR. Non-reported patients were older and had a considerable poorer survival than those included in the SCR. Inclusion of DCI cases decreased one-year lung cancer overall survival from 45% to 41%. The corresponding decrease for pancreatic cancer was five percentage points, from 29% to 24%. Conclusions Lung and pancreatic cancers are underreported to the SCR yielding too low incidence rates and upward biased survival estimates. We conclude that implementation of systematic death certificate processing with trace-back is feasible also within the Swedish system with regionalized cancer reporting. Verifying registrability by use of information in the Patient Register provided a good approximation of “corrected” survival estimates based on chart review.

Effect of Net‐Gun Capture on Survival of Mule Deer

ABSTRACTCapture techniques to deploy radio‐collars often risk mortality and injury to the animal. Capture‐induced mortality can affect population sizes but also introduces bias in survival estimates based on data from captured animals. In recent years, a large‐scale research and monitoring project in Utah, USA, has involved capturing and radio‐collaring hundreds of mule deer (Odocoileus hemionus), a species of great interest in large parts of North America. Our objective was to investigate how the survival rates of these mule deer were affected by capture and handling. During winters of 2014–2018, an experienced capture crew net‐gunned and fitted 1,805 animals with global positioning system (GPS)‐collars. We estimated survival rates during the first 6 weeks after capture using Cox proportional hazard regression, and compared the survival rates of animals that were captured in a particular year to those of animals that were not captured but fitted with a GPS‐collar in a previous year. We used a model selection framework to evaluate how long survival rates of captured animals were different from those of animals that were not captured. Our results indicated that weekly survival rates of captured animals were 0.985 ± 0.003 (SE), 0.988 ± 0.002, and 0.990 ± 0.001 in weeks 1, 2 and 3, respectively. Weekly survival rates of captured deer during weeks 4–6 were 0.993 ± 0.001, the same as those of deer that were not captured at the same time. Furthermore, post‐capture survival rates were positively influenced by body size and negatively influenced by age. We conclude that the mortality resulting from helicopter capture was low but recommend comparing newly captured and previously captured individuals to examine what proportion of observed mortality is likely capture‐related. © 2020 The Wildlife Society.

Assessing Selection Bias in Non-Experimental Estimates of the Returns to Workplace Training

We assess selection bias in estimated returns to workplace training by exploiting a field experiment with random assignment of workers to a one-week training program. We compare experimental estimates of this program with non-experimental estimates that are estimated by using a sample of agents who were selected by management not to participate in the experiment. Our results show that non-experimental estimates are biased, yielding returns about twice as large as the causal effect. When controlling for pre-treatment performance or individual fixed effects, only about one tenth of this bias remains and is even further reduced when applying common support restrictions.

Features of survival analysis on patients on the «waiting list» for kidney transplantation

Survival analysis is one of the most common methods of statistical analysis in medicine. The statistical analysis of the transplantation (or death) probability dependent on the waiting time on the "waiting list" is a rare case when the survival analysis is used to estimate the time before the event rather than to indirectly assess the risks. However, for an assessment to be adequate, the reason for censoringmust be independent of the outcome of interest. Patients on the waiting list are not only at risk of dying, they can be excluded from the waiting list due to deterioration of the comorbid background or as a result of kidney transplantation. Kaplan – Meier, Nelson – Aalen estimates, as well as a cause-specific Cox proportional hazards regression model, are consciously biased estimates of survival in the presence of competing risks. Since competing events are censored, it is impossible to directly assess the impact of covariates on their frequency, because there is no direct relationship between the regression coefficients and the intensity of these events. The determination of the median waiting time on the basis of such analysis generates a selection bias, which inevitably leads to a biased assessment. Thus, in presence of competing risks, these methods allow us to investigate the features of cause-and-effect relationships, but do not allow us to make a prediction of the individual probability of a particular event based on the value of its covariates. In the regression model of competing risks, the regression coefficients are monotonically related to the cumulative incidence function and the competing events have a direct impact on the regression coefficients. Its significant advantage is the additive nature of the cumulative incidence functions of all possible events. In the study of etiological associations, it is better to use Cox regression model, which allows to estimate the size of the effect of various factors. The regression model of competing risks, in turn, has a greater prognostic value and allows to estimate the probability of a specific outcome within a certain time in a single patient.

Indices of non-ignorable selection bias for proportions estimated from non-probability samples.

Rising costs of survey data collection and declining response rates have caused researchers to turn to non-probability samples to make descriptive statements about populations. However, unlike probability samples, non-probability samples may produce severely biased descriptive estimates due to selection bias. The paper develops and evaluates a simple model-based index of the potential selection bias in estimates of population proportions due to non-ignorable selection mechanisms. The index depends on an inestimable parameter ranging from 0 to 1 that captures the amount of deviation from selection at random and is thus well suited to a sensitivity analysis. We describe modified maximum likelihood and Bayesian estimation approaches and provide new and easy-to-use R functions for their implementation. We use simulation studies to evaluate the ability of the proposed index to reflect selection bias in non-probability samples and show how the index outperforms a previously proposed index that relies on an underlying normality assumption. We demonstrate the use of the index in practice with real data from the National Survey of Family Growth.

Occurrence, fate, and confounding influence of ghost passive integrated transponder tags in an intensively monitored watershed

Over the last three decades, passive integrated transponder (PIT) tags have been widely used to study fish populations. Interpretation of PIT tag detections, however, can be confounded by the presence of ghost tags, tags liberated when a fish dies. We used a combination of mobile antenna surveys, stationary antenna detections, and multistate mark–recapture modeling to assess the abundance and fate of ghost tags in a coastal California watershed. Accumulation of ghost tags from released hatchery-origin coho salmon (Oncorhynchus kisutch) smolts was substantial during California’s recent drought, with 2224 ghost tags identified during mobile reader surveys. Between surveys, PIT tags moved downstream a median distance of 346 m and a maximum distance of 1982 m. Stationary antenna array detections indicated that these movements occurred during high-flow events, concurrent with live fish movement. The multistate model estimated that, during winter, approximately 40% of tags were buried in the substrate beyond the read range of mobile readers. Failure to account for transport and burial dynamics of ghost tags can lead to biased estimates of fish abundance, survival, and movement.

Bias in survival estimates created by a requirement for consent to enter a clinical breast cancer registry

BackgroundA requirement for consent for inclusion may bias the results from a clinical registry. This study gives a direct measure of this bias, based on a population-based clinical breast cancer registry where the requirement for consent was removed after further ethical review and data could be re-analysed. MethodsIn Auckland, New Zealand, the population-based clinical breast cancer registry required written patient consent for inclusion from 2000-2012. A subsequent ethical review removed this requirement and allowed an analysis of consented and non-consented patients. Kaplan-Meier survival to 10 years (mean follow-up 5.1 years, maximum 13.9 years), demographic and clinical characteristics were compared. Of 9244 women with invasive cancer, 926 (10.4%) were not consented, and of 1642 women with ductal carcinoma in situ, 245 (14.9%) were not consented. ResultsSurvival was much higher for consenting patients; invasive cancer, 5 year survival 83.2% (95% confidence limits 82.2–84.1%) for consenting patients, 57.1% (53.0–60.9%) for non-consenting, and 80.8% in all patients. Analyses based only on consenting patients overestimate survival in all patients by around 2% at 2, 5, and 10 years. Non-consented patients were older, more often of Pacific ethnicity, had fewer screen-detected cancers, and more often had metastatic disease; they less frequently had primary surgery or systemic treatments. ConclusionData from a registry requiring active consent gives an upward bias in survival results, as non-consenting patients have more extensive disease, less treatment, and lower survival. To give unbiased results active consent should be not required in a clinical cancer registry.

Selection bias in estimation of peer effects in product adoption

It is well recognized that consistent estimation of peer effects faces formidable identification challenges. The confounding factors the literature usually focuses on are endogenous group formation, correlated unobservables and simultaneity. In this paper, I show that another significant source of bias arises when a researcher examines peer effects in product adoption. When people differ in their valuations of a product, people with higher valuations tend to adopt earlier so that only those with increasingly lower valuations comprise the set of potential adopters as time progresses. Such an endogenous attrition over time, if not correctly accounted for, will lead to inconsistent estimates of peer effects. I present simulations to numerically demonstrate the presence and extent of such a selection bias. I also propose a simple solution to remove the bias and examine its performance.